Start by measuring cost per successful task, then reduce repeated context, unnecessary output, and oversized model calls. Validate every change against representative evals.

Measure the workflow, not just the invoice

Separate input tokens, output tokens, retries, and failed tasks by workflow. A lower price per call is not a saving if the task needs more calls or creates more manual review.

Test one change at a time

Trim unused context, set appropriate output limits, and test routing simpler tasks to a smaller model. Compare output quality, latency, and total cost on the same set of real examples. Keep a rollback path for changes that fail.